Avionics Certification Wasn't Built to Evaluate Judgment

Avionics Certification Wasn’t Built to Evaluate Judgment

The hard problem in AI-enabled avionics is not primarily a headcount problem. It is a category problem, and it will persist even if regulators hire their way out of the queue.

For decades, avionics software assurance has rested on a straightforward premise. Code executes deterministic logic, and you can trace every requirement to a test case that proves the logic does what it says. DO-178C, the baseline standard behind commercial avionics software approval, was built on that premise, according to an industry certification overview of the process DO-178 certification overview. That premise is now under pressure from systems that do not execute fixed logic so much as produce a judgment.

The clearest signal of that shift is also the one that needs the most scrutiny. NASA’s Phase II contract with AstraQua, as the company describes it, names “health-informed reasoning” for autonomous platforms as the capability under development, with verification and validation framed as the gating requirement for systems flying over populated areas AstraQua. That description comes from a vendor announcement about a small early-stage award, not an independently verified engineering milestone, and it should be read as a market signal of where contracting language is heading rather than evidence that the V&V problem has been solved. Treated that way, it is still useful. It tells program leads what capability regulators and primes are being asked to fund, even before anyone has a settled way to certify it.

On the unmanned side, the same shift is underway for a more durable reason. Edge AI is moving drone systems from simple object detection to contextual understanding of a scene, which means the system being certified is no longer a sensor pipeline but an inference the aircraft makes about its environment Military Embedded Systems. Government postings for AI and machine learning avionics engineers now reflect that reality directly, specifying lab, SIL/HIL, subsystem, and platform-level verification as core deliverables, which is a good proxy for where engineering hours are actually going on these programs Careers in Government.

None of that work has a settled methodology yet. You can trace a line of code to a requirement. It is far less clear how you trace a contextual judgment to one, or what evidence satisfies a certifying authority that the judgment will hold across cases it was never explicitly trained on. That open question exists independent of regulator staffing levels.

It is also compounded by them. The Aerospace, Security and Defence Industries Association of Europe has flagged that EASA operates its certification function with roughly a quarter of the full-time staff the FAA runs, a gap it is now presenting to policymakers as a structural risk rather than a budget line item ASD Europe. A thinner regulatory bench does not just mean slower throughput on familiar filings. It means less capacity to develop new assurance methodology for a class of system that does not yet have one, at the exact moment that methodology is most needed.

Program leads sponsoring AI-enabled avionics work should separate these two problems rather than collapsing them into a scheduling estimate. The first question is methodological: what counts as sufficient evidence that a reasoning system’s judgment is sound, and who inside the program is building that evidence model now, before a filing forces the question. The second is logistical: which regulator’s queue you are actually in, and whether that regulator currently has the bandwidth to evaluate novel evidence rather than process a familiar checklist. Budgeting for the second without first solving the first is how a certification timeline quietly turns into a certification standstill.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The core argument—that AI-enabled avionics face a category problem (certifying judgment vs. deterministic logic) distinct from a staffing problem—is coherent and well-structured, though the AstraQua e
Source & Claim Verification	In-house seat · hosted fallback (qwen-2.5-72b-instruct)	cleared. The draft accurately cites and supports most claims, but the claim about the number of FTEs at EASA and FAA is slightly off in the cited source.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately reflects DO-178C’s limitations for AI judgment evaluation and aligns with current aerospace regulatory challenges without overstepping other frameworks.
Technical Accuracy	Llama	cleared. The article accurately identifies the challenge of certifying AI-enabled avionics systems under current standards like DO-178C, highlighting the methodological gap in verifying judgment-based systems.
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and flags vendor-driven language and market signals, distinguishing them from independently verified engineering milestones or settled methodologies.
Novelty & Non-Duplication	Grok	held. Restates the EASA staffing-gap + AI-avionics assurance/methodology thesis already briefed repeatedly (esp. 2026-11-02 Where Aerospace AI Is Already Flying, 2026-10-31 Aviation’s AI Rules/Regulator Bot
Validation	DeepSeek	cleared. The central claim is factually supportable because DO-178C is deterministic and requirements-based while AI/ML judgment certification lacks settled methods, but the draft’s sourcing is weak, including

Sources cited: 11. Validation challenges: 0. Learn how these briefings are written and verified.